Causal DAG(Directed Acyclic Graph) usually lies in a 2D plane without distinguishing correlation changes and causal effects. Also, the causal effect is often approximately estimated by averaging the population's correlation changes. Now, AI(Artificial Intelligence) enables much larger-scale structural modeling, whose complex hidden confoundings make the approximation errors no longer ignorable but can snowball to considerable population-level Causal Representation Bias. Such bias has caused significant problems: ungeneralizable causal models, unrevealed individual-level features, not utilizable causal knowledge in DL(Deep Learning), etc. In short, DAG must be redefined to enable a new framework for causal AI. Observational time series can only reflect correlation changes in statistics. But the DL-based autoencoder can represent them as individual-level feature changes in latent space to reflect causal effects. In this paper, we introduce the redefined do-DAG concept and propose Causal Representation Learning (CRL) framework as the generic solution, along with a novel architecture to realize CRL and experimentally verify its feasibility.

翻译：因果DAG（有向无环图）通常处于二维平面中，无法区分相关性变化与因果效应。此外，因果效应常通过总体相关性变化的平均估计进行近似计算。当前，人工智能技术实现了更大规模的结构建模，其复杂隐变量混杂使得近似误差不再可忽略，反而可能累积形成显著的总体级因果表示偏差。此类偏差已引发严重问题：因果模型缺乏泛化能力、个体级特征无法被揭示、深度学习中的因果知识难以利用等。简言之，DAG必须重新定义以构建面向因果AI的新框架。观测时序数据仅能反映统计层面的相关性变化，而基于深度学习的自编码器可将其表征为隐空间中的个体级特征变化，从而体现因果效应。本文引入重定义的do-DAG概念，提出因果表示学习框架作为通用解决方案，并设计新型架构实现因果表示学习，通过实验验证其可行性。